Research · ASDICOn the phone everything happens in seconds: half of answered calls last ten seconds or less. Off-the-shelf models were built for read speech, in English. We are building our own, for phone conversation, in French, Spanish and English, trained and hosted in Europe.
In development · research project 2027-2028
ASDIC is the ancestor of sonar, developed during the First World War with the decisive contribution of French physicist Paul Langevin. It sends a sound, listens to the echo and decides within seconds whether it hears a real target or just noise.
Our model does the same thing on the phone: it listens, tells a human from a voicemail at once, spots a real intent in the middle of the conversation, and raises its hand at the right moment.
The model, its weights and its corpus belong to Nodical. It is trained on European computing power and hosted in the European Union. No single vendor can switch it off.
From the pick-up, ASDIC says who the agent is talking to. Today the decision lands after eight seconds (median) on our campaigns. The goal: under two seconds.
Interested, not interested, call back later, wrong number, asks for a human, does not want to be contacted again: the intent is recognised while the person speaks, along with the answers to eligibility questions.
Neither cutting them off nor leaving a gap. Waiting for silence is not enough: it is the weak point of every voice agent, and the heart of our work.
When an adviser should take over, ASDIC says so, with a two-line summary so nobody arrives cold.
French, Spanish, English: the same model, the same performance, without rebuilding one model per language.
Read from Nodical's production database, without reading the content of conversations. These are our starting points: targets will be published with the measurements, not before.
ASDIC listens and understands. It never produces a voice and never learns to imitate one.
It helps the agent lead the conversation. It decides nothing about the person called.
Calls used for training are pseudonymised: names, numbers and addresses are removed before any annotation. Anyone can object by a simple email.
A corpus of pseudonymised calls, annotated in French, then Spanish and English. None exists in open access for French telephone speech.
Off-the-shelf models measured on this corpus: the quantified starting point.
Design and training of the first four functions.
Spanish and English, accents included.
Integration into Nodical and trials in real conditions, on live campaigns.
Final measurements, published; protection of the model; roll-out.
With the support of a public research laboratory in the Paris region specialised in speech processing, currently being selected.
